OpenLipid: a large language model workflow for targeted analysis of DIA mass spectrometry data in lipidomics

OpenLipid applies LLMs to chromatographic evidence evaluation in targeted DIA lipidomics, identifying candidate peak groups directly from XICs and producing peak/no-peak decisions with peak boundaries and supporting product ions, without lipidomics-specific training or fine-tuning. The prior ChatDIA study demonstrated LLM-based XIC evaluation in targeted DIA proteomics; this work extends that approach to lipidomics and systematically assesses its ability to identify target lipid signals. Benchmarked against manual annotations on four datasets (SRM 1950 plasma and mouse feces, positive and negative ionization modes), with each lipid target analyzed independently and three replicate runs (R0-R2).

At 5% FDR, OpenLipid (LLM Majority ensemble) accepted 110, 28, 130, and 84 identifications, corresponding to 55.3%, 19.0%, 71.8%, and 31.8% of library targets, comparable overall to DIAMetAlyzer (57.8%, 19.7%, 89.0%, 17.8%) and substantially higher than untargeted MS-DIAL DIA (12.6%, 0.0%, 47.5%, 0.0%). Provides a first FDR-coverage comparison of an LLM workflow against targeted (DIAMetAlyzer) and untargeted (MS-DIAL DIA) methods on the same assay-library targets in lipidomics. Manual annotations served as ground truth; FDR was the proportion of accepted identifications that were incorrect and coverage the proportion of annotated library targets accepted; coverage was consistently higher in positive than negative ionization mode.

At the stricter 1% FDR threshold, OpenLipid achieved coverage of 35.7%, 12.2%, 39.2%, and 22.7%, higher than DIAMetAlyzer (33.7%, 4.8%, 70.7%, 15.5%) in three of four datasets, with no incorrect candidates observed among accepted identifications. Indicates that LLM-derived scores retain candidate-ranking ability in the low-FDR regime, where MS-DIAL DIA had no operating point satisfying the threshold in the negative-ionization datasets. FDR-coverage curves and precision-recall analysis; OpenLipid average precision was 0.903, 0.851, 0.924, and 0.900 versus 0.865, 0.732, 0.960, and 0.738 for DIAMetAlyzer and 0.451, 0.232, 0.544, and 0.325 for MS-DIAL DIA.

Individual LLM-reported chromatographic features (retention-time agreement, ion co-elution, peak shape, fragment-ion pattern consistency, and number of supporting product ions) supported supervised classifiers distinguishing correct from incorrect candidate peak groups and retained predictive information in bidirectional transfer between plasma and feces. Shows that LLM outputs can serve not only as final decisions but as feature inputs for downstream statistical modeling, with candidate-ranking information transferring across sample types. Repeated stratified five-fold cross-validation grouped by lipid target with 10 repeats; median average precision was 0.838-0.912 for LLM-derived features and 0.734-0.865 for DIAMetAlyzer-derived features, all above candidate-prevalence baselines of 0.117-0.270; external validation median average precision was 0.828-0.915 for LLM features.

The work targets settings where assay libraries are built from DDA results and predefined lipid targets are evaluated in DIA data, applied to human plasma and mouse feces in positive and negative ionization modes acquired on the SCIEX ZenoTOF 7600 platform; its value lies in offering lipidomics researchers FDR-controlled automated candidate peak evaluation with human-readable rationales, and in serving as a source of features for downstream statistical modeling.

Evaluation used four datasets from a single instrument platform, so broader applicability across laboratories, acquisition settings, and chromatographic conditions remains to be established; the FDR-controlled comparison with MS-DIAL DIA was restricted to assay-library targets; the workflow relies on MS-DIAL for DDA-based identification and library construction, and MS-DIAL does not provide specific fragment identities in its lipid MS/MS annotations, limiting the structural information available to the LLM when evaluating fragment-ion evidence; score cutoffs selected at nominal FDR levels may need adjustment for new datasets; and further evaluation with additional runs and repeated ensembles would help determine the number of runs needed and quantify remaining variability.